Abstract
This paper presents a magnetic-flux-based encoder for BLDC drives that maintains high accuracy under rotor eccentricity and dynamic transients. Conventional Hall-sensor-based angle estimators rely on ideal sinusoidal flux assumptions and degrade in the presence of air-gap distortion or misalignment. To overcome these limitations, a nonlinear autoregressive network with exogenous inputs (NARXNet) is proposed as a temporal neural observer that learns the nonlinear, time-dependent mapping between measured flux densities and the true electrical rotor angle. A physics-informed data augmentation framework combines experimentally measured magnetic flux maps with dynamic simulation to generate diverse training scenarios at low and variable speeds. Validation demonstrates mean angular errors below 2°, 95th-percentile errors under 5°, and negligible drift, with enhanced resilience to eccentric displacement and acceleration transients compared to classical methods. The proposed approach provides a compact, data-driven sensing solution for robust, encoderless electric drive control.